张洪交, 张存喜, 王瑞, 王可, 乔倩. 基于图像处理和改进DenseNet网络的小黄鱼新鲜度识别[J]. 南方水产科学, 2024, 20(3): 133-142. DOI: 10.12131/20230241
引用本文: 张洪交, 张存喜, 王瑞, 王可, 乔倩. 基于图像处理和改进DenseNet网络的小黄鱼新鲜度识别[J]. 南方水产科学, 2024, 20(3): 133-142. DOI: 10.12131/20230241
ZHANG Hongjiao, ZHANG Cunxi, WANG Rui, WANG Ke, QIAO Qian. Freshness recognition of small yellow croaker based on image processing and improved DenseNet network[J]. South China Fisheries Science, 2024, 20(3): 133-142. DOI: 10.12131/20230241
Citation: ZHANG Hongjiao, ZHANG Cunxi, WANG Rui, WANG Ke, QIAO Qian. Freshness recognition of small yellow croaker based on image processing and improved DenseNet network[J]. South China Fisheries Science, 2024, 20(3): 133-142. DOI: 10.12131/20230241

基于图像处理和改进DenseNet网络的小黄鱼新鲜度识别

Freshness recognition of small yellow croaker based on image processing and improved DenseNet network

  • 摘要: 传统水产品新鲜度检测方法存在对样本破坏较大、操作步骤繁琐、检测准确率及效率较低等一系列问题。针对上述问题,以小黄鱼(Larimichthys polyactis)新鲜度高效、准确识别为目标,提出了一种基于改进DenseNet网络的小黄鱼新鲜度识别模型。首先,在DenseNet网络结构中的每个Denseblock模块引入SENet注意力机制模块,实现特征通道特征重标定,加强网络对当前有益特征的提取,摒弃无作用的特征。其次,改进卷积层的第一层,增加网络的非线性能力和特征表达能力。为防止训练过程中出现梯度消失的现象,用PReLU激活函数代替原网络的ReLU激活函数。最后,与原DenseNet网络模型及其他经典神经网络模型进行对比实验。结果表明,构建的基于迁移学习的FishNet模型在自建的小黄鱼新鲜度数据集上识别准确率达91.53%,模型具有较高的识别准确率和较强的鲁棒性,解决了水产品新鲜度检测高效和精准识别问题,也为开发智能新鲜度识别系统提供了参考。

     

    Abstract: Traditional freshness detection methods for aquatic products have problems such as great sample damage, trivial operation steps, low detection accuracy and efficiency. To solve these problems, in order to efficiently and accurately identify the freshness of small yellow croaker (Larimichthys polyactis), we proposed a freshness recognition model based on an improved DenseNet network. Firstly, we introduced the SENet attention mechanism module into each dense block module in the DenseNet network structure to achieve feature channel feature recalibration, enhance the network's extraction of current beneficial features, and eliminate irrelevant features. Secondly, we improved the first layer of the convolutional layer to enhance the network's non-linear ability and feature representation ability. To prevent the phenomenon of gradient vanishing during the training process, we used the PReLU activation function instead of the ReLU activation function of the original network. Finally, we conducted comparative experiments with the DenseNet network model and other classic neural network models. The experimental results show that the FishNet model based on transfer learning constructed in this paper has a recognition accuracy of 91.53% on the built L. polyactis freshness dataset. The model has high recognition accuracy and strong robustness, achieving efficient and accurate recognition of aquatic product freshness detection, and providing references for the development of intelligent freshness recognition systems.

     

/

返回文章
返回